Sungwoo Kim
distribution-201229
Diff: main.cpp
- Revision:
- 177:8e9cf31d63f4
- Parent:
- 176:589ea3edcf3c
- Child:
- 178:1074553d2f6f
--- a/main.cpp Tue Nov 24 10:16:10 2020 +0000
+++ b/main.cpp Tue Nov 24 13:04:54 2020 +0000
@@ -1,4 +1,4 @@
-//201124_4
+//201124_5
#include "mbed.h"
#include "FastPWM.h"
#include "INIT_HW.h"
@@ -489,38 +489,35 @@
void update_Critic_Networks(float (*arr)[num_input_RL])
{
- float gradient_rate = 0.001f;
+ float gradient_rate = 0.1f;
float G_hc1[num_input_RL][num_hidden_unit1] = {0.0f};
- float d_V_d_hc1[batch_size][num_input_RL][num_hidden_unit1] = {0.0f}; ////////////////1
- float G_bc1[num_hidden_unit1] = {0.0f};
- float d_V_d_bc1[batch_size][num_hidden_unit1] = {0.0f}; ////////////////2
+ float G_bc1[num_hidden_unit1] = {0.0f};
for (int index2 = 0; index2 < num_hidden_unit1; index2++) {
for (int index1 = 0; index1 < num_input_RL; index1++) {
for (int n=0; n<batch_size; n++) {
+ float d_V_d_hc1 = 0.0f;
for(int k=0; k<num_hidden_unit2; k++) {
if (hxh_c_sum_array[n][k] >= 0) {
if (hx_c_sum_array[n][index2] > 0) {
- //G_hc1[index1][index2] = G_hc1[index1][index2] + arr[n][index1]*hc2_temp[index2][k]*hc3_temp[k]; //////////////////////3
- d_V_d_hc1[n][index1][index2] = d_V_d_hc1[n][index1][index2] + arr[n][index1]*hc2_temp[index2][k]*hc3_temp[k]; //////////////////////4
+ d_V_d_hc1 = d_V_d_hc1 + arr[n][index1]*hc2_temp[index2][k]*hc3_temp[k];
}
}
}
- G_hc1[index1][index2] = G_hc1[index1][index2] + 2.0f*(return_G[n]-V[n])*(-d_V_d_hc1[n][index1][index2]); /////////////////////5
+ G_hc1[index1][index2] = G_hc1[index1][index2] + 2.0f*(return_G[n]-V[n])*(-d_V_d_hc1);
}
G_hc1[index1][index2] = G_hc1[index1][index2] / batch_size;
//hc1_temp[index1][index2] = hc1_temp[index1][index2] - gradient_rate * G_hc1[index1][index2];
}
for (int n=0; n<batch_size; n++) {
+ float d_V_d_bc1 = 0.0f;
for(int k=0; k<num_hidden_unit2; k++) {
if (hxh_c_sum_array[n][k] >= 0) {
if (hx_c_sum_array[n][index2] > 0) {
- //G_bc1[index2] = G_bc1[index2] + hc2_temp[index2][k]*hc3_temp[k]; //////////////////6
- d_V_d_bc1[n][index2] = d_V_d_bc1[n][index2] + hc2_temp[index2][k]*hc3_temp[k]; //////////////////7
+ d_V_d_bc1 = d_V_d_bc1 + hc2_temp[index2][k]*hc3_temp[k];
}
-
}
}
- G_bc1[index2] = G_bc1[index2] + 2.0f*(return_G[n]-V[n])*(-d_V_d_bc1[n][index2]); /////////////////////8
+ G_bc1[index2] = G_bc1[index2] + 2.0f*(return_G[n]-V[n])*(-d_V_d_bc1);
}
G_bc1[index2] = G_bc1[index2] / batch_size;
//bc1_temp[index2] = bc1_temp[index2] - gradient_rate * G_bc1[index2];
@@ -528,54 +525,50 @@
float G_hc2[num_hidden_unit1][num_hidden_unit2] = {0.0f};
- float d_V_d_hc2[batch_size][num_hidden_unit1][num_hidden_unit2] = {0.0f};
float G_bc2[num_hidden_unit2] = {0.0f};
- float d_V_d_bc2[batch_size][num_hidden_unit2] = {0.0f};
for (int index2 = 0; index2 < num_hidden_unit2; index2++) {
for (int index1 = 0; index1 < num_hidden_unit1; index1++) {
for (int n=0; n<batch_size; n++) {
+ float d_V_d_hc2 = 0.0f;
if (hxh_c_sum_array[n][index2] >= 0) {
if (hx_c_sum_array[n][index1] > 0) {
- //G_hc2[index1][index2] = G_hc2[index1][index2] + hx_c_sum_array[n][index1]*hc3_temp[index2];
- d_V_d_hc2[n][index1][index2] = hx_c_sum_array[n][index1]*hc3_temp[index2];
+ d_V_d_hc2 = hx_c_sum_array[n][index1]*hc3_temp[index2];
}
}
- G_hc2[index1][index2] = G_hc2[index1][index2] + 2.0f*(return_G[n]-V[n])*(-d_V_d_hc2[n][index1][index2]);
+ G_hc2[index1][index2] = G_hc2[index1][index2] + 2.0f*(return_G[n]-V[n])*(-d_V_d_hc2);
}
G_hc2[index1][index2] = G_hc2[index1][index2] / batch_size;
//hc2_temp[index1][index2] = hc2_temp[index1][index2] - gradient_rate * G_hc2[index1][index2];
}
for (int n=0; n<batch_size; n++) {
+ float d_V_d_bc2 = 0.0f;
if (hxh_c_sum_array[n][index2] >= 0) {
- //G_bc2[index2] = G_bc2[index2] + hc3_temp[index2];
- d_V_d_bc2[n][index2] = hc3_temp[index2];
+ d_V_d_bc2 = hc3_temp[index2];
}
- G_bc2[index2] = G_bc2[index2] + 2.0f*(return_G[n]-V[n])*(-d_V_d_bc2[n][index2]);
+ G_bc2[index2] = G_bc2[index2] + 2.0f*(return_G[n]-V[n])*(-d_V_d_bc2);
}
G_bc2[index2] = G_bc2[index2] / batch_size;
//bc2_temp[index2] = bc2_temp[index2] - gradient_rate * G_bc2[index2];
}
float G_hc3[num_hidden_unit2]= {0.0f};
- float d_V_d_hc3[batch_size][num_hidden_unit2] = {0.0f};
float G_bc3 = 0.0f;
- float d_V_d_bc3[batch_size] = {0.0f};
for (int index2 = 0; index2 < 1; index2++) {
for (int index1 = 0; index1 < num_hidden_unit2; index1++) {
for (int n=0; n<batch_size; n++) {
+ float d_V_d_hc3 = 0.0f;
if (hxh_c_sum_array[n][index1] >= 0) {
- //G_hc3[index1] = G_hc3[index1] + hxh_c_sum_array[n][index1];
- d_V_d_hc3[n][index1] = d_V_d_hc3[n][index1] + hxh_c_sum_array[n][index1];
+ d_V_d_hc3 = d_V_d_hc3 + hxh_c_sum_array[n][index1];
}
- G_hc3[index1] = G_hc3[index1] + 2.0f*(return_G[n]-V[n])*(-d_V_d_hc3[n][index1]);
+ G_hc3[index1] = G_hc3[index1] + 2.0f*(return_G[n]-V[n])*(-d_V_d_hc3);
}
G_hc3[index1] = G_hc3[index1] / batch_size;
//hc3_temp[index1] = hc3_temp[index1] - gradient_rate * G_hc3[index1];
}
for (int n=0; n<batch_size; n++) {
- //G_bc2[index2] = G_bc2[index2] + 1.0f;
- d_V_d_bc3[n] = 1.0f;
- G_bc3 = G_bc3 + 2.0f*(return_G[n]-V[n])*(-d_V_d_bc3[n]);
+ float d_V_d_bc3 = 0.0f;
+ d_V_d_bc3 = 1.0f;
+ G_bc3 = G_bc3 + 2.0f*(return_G[n]-V[n])*(-d_V_d_bc3);
}
G_bc3 = G_bc3 / batch_size;
//bc3_temp = bc3_temp - gradient_rate * G_bc3;
@@ -600,39 +593,36 @@
}
bc3_temp = bc3_temp - gradient_rate * G_bc3;
}
-
}
///////////////////////////Softplus//////////////////////////////////
void update_Actor_Networks(float (*arr)[num_input_RL])
{
- float gradient_rate = 0.001f;
+ float gradient_rate = 0.1f;
float G_ha1[num_input_RL][num_hidden_unit1] = {0.0f};
float G_ba1[num_hidden_unit1] = {0.0f};
- float d_x_d_ha1[batch_size][num_input_RL][num_hidden_unit1] = {0.0f};
- float d_x_d_ba1[batch_size][num_hidden_unit1] = {0.0f};
- float d_y_d_ha1[batch_size][num_input_RL][num_hidden_unit1] = {0.0f};
- float d_y_d_ba1[batch_size][num_hidden_unit1] = {0.0f};
for (int index2 = 0; index2 < num_hidden_unit1; index2++) {
for (int index1 = 0; index1 < num_input_RL; index1++) {
for (int n=0; n<batch_size; n++) {
+ float d_x_d_ha1 = 0.0f;
+ float d_y_d_ha1 = 0.0f;
if((advantage[n] >= 0.0f && ratio[n] >= 1.0f + epsilon) || (advantage[n] < 0.0f && ratio[n] < 1.0f - epsilon)) {
G_ha1[index1][index2] = G_ha1[index1][index2];
} else {
for(int k=0; k<num_hidden_unit2; k++) {
if (hxh_a_sum_array[n][k] >= 0) {
if (hx_a_sum_array[n][index2] > 0) {
- d_x_d_ha1[n][index1][index2] = d_x_d_ha1[n][index1][index2] + arr[n][index1]*ha2_temp[index2][k]*ha3_temp[k][0];
- d_y_d_ha1[n][index1][index2] = d_y_d_ha1[n][index1][index2] + arr[n][index1]*ha2_temp[index2][k]*ha3_temp[k][1];
+ d_x_d_ha1 = d_x_d_ha1 + arr[n][index1]*ha2_temp[index2][k]*ha3_temp[k][0];
+ d_y_d_ha1 = d_y_d_ha1 + arr[n][index1]*ha2_temp[index2][k]*ha3_temp[k][1];
}
}
}
float d_mean_d_ha1 = 0.0f;
float d_dev_d_ha1 = 0.0f;
- d_mean_d_ha1 = exp(hxhh_a_sum_array[n][0])/(1.0f+exp(hxhh_a_sum_array[n][0]))*d_x_d_ha1[n][index1][index2];
- d_dev_d_ha1 = exp(hxhh_a_sum_array[n][1])/(1.0f+exp(hxhh_a_sum_array[n][1]))*d_y_d_ha1[n][index1][index2];
+ d_mean_d_ha1 = exp(hxhh_a_sum_array[n][0])/(1.0f+exp(hxhh_a_sum_array[n][0]))*d_x_d_ha1;
+ d_dev_d_ha1 = exp(hxhh_a_sum_array[n][1])/(1.0f+exp(hxhh_a_sum_array[n][1]))*d_y_d_ha1;
G_ha1[index1][index2] = G_ha1[index1][index2] + advantage[n]/pi_old[n]*(d_mean_d_ha1*Grad_Normal_Dist_Mean(mean_array[n],deviation_array[n],action_array[n])+d_dev_d_ha1*Grad_Normal_Dist_Deviation(mean_array[n],deviation_array[n],action_array[n]));
}
@@ -642,21 +632,23 @@
}
for (int n=0; n<batch_size; n++) {
+ float d_x_d_ba1 = 0.0f;
+ float d_y_d_ba1 = 0.0f;
if((advantage[n] >= 0.0f && ratio[n] >= 1.0f + epsilon) || (advantage[n] < 0.0f && ratio[n] < 1.0f - epsilon)) {
G_ba1[index2] = G_ba1[index2];
} else {
for(int k=0; k<num_hidden_unit2; k++) {
if (hxh_a_sum_array[n][k] >= 0) {
if (hx_a_sum_array[n][index2] > 0) {
- d_x_d_ba1[n][index2] = d_x_d_ba1[n][index2] + ha2_temp[index2][k]*ha3_temp[k][0];
- d_y_d_ba1[n][index2] = d_y_d_ba1[n][index2] + ha2_temp[index2][k]*ha3_temp[k][1];
+ d_x_d_ba1 = d_x_d_ba1 + ha2_temp[index2][k]*ha3_temp[k][0];
+ d_y_d_ba1 = d_y_d_ba1 + ha2_temp[index2][k]*ha3_temp[k][1];
}
}
}
float d_mean_d_ba1 = 0.0f;
float d_dev_d_ba1 = 0.0f;
- d_mean_d_ba1 = exp(hxhh_a_sum_array[n][0])/(1.0f+exp(hxhh_a_sum_array[n][0]))*d_x_d_ba1[n][index2];
- d_dev_d_ba1 = exp(hxhh_a_sum_array[n][1])/(1.0f+exp(hxhh_a_sum_array[n][1]))*d_y_d_ba1[n][index2];
+ d_mean_d_ba1 = exp(hxhh_a_sum_array[n][0])/(1.0f+exp(hxhh_a_sum_array[n][0]))*d_x_d_ba1;
+ d_dev_d_ba1 = exp(hxhh_a_sum_array[n][1])/(1.0f+exp(hxhh_a_sum_array[n][1]))*d_y_d_ba1;
G_ba1[index2] = G_ba1[index2] + advantage[n]/pi_old[n]*(d_mean_d_ba1*Grad_Normal_Dist_Mean(mean_array[n],deviation_array[n],action_array[n])+d_dev_d_ba1*Grad_Normal_Dist_Deviation(mean_array[n],deviation_array[n],action_array[n]));
}
@@ -667,28 +659,26 @@
float G_ha2[num_hidden_unit1][num_hidden_unit2] = {0.0f};
float G_ba2[num_hidden_unit2] = {0.0f};
- float d_x_d_ha2[batch_size][num_hidden_unit1][num_hidden_unit2] = {0.0f};
- float d_x_d_ba2[batch_size][num_hidden_unit2] = {0.0f};
- float d_y_d_ha2[batch_size][num_hidden_unit1][num_hidden_unit2] = {0.0f};
- float d_y_d_ba2[batch_size][num_hidden_unit2] = {0.0f};
for (int index2 = 0; index2 < num_hidden_unit2; index2++) {
for (int index1 = 0; index1 < num_hidden_unit1; index1++) {
for (int n=0; n<batch_size; n++) {
+ float d_x_d_ha2 = 0.0f;
+ float d_y_d_ha2 = 0.0f;
if((advantage[n] >= 0.0f && ratio[n] >= 1.0f + epsilon) || (advantage[n] < 0.0f && ratio[n] < 1.0f - epsilon)) {
G_ha2[index1][index2] = G_ha2[index1][index2];
} else {
if (hxh_a_sum_array[n][index2] >= 0) {
if (hx_a_sum_array[n][index1] > 0) {
- d_x_d_ha2[n][index1][index2] = hx_a_sum_array[n][index1]*ha3_temp[index2][0];
- d_y_d_ha2[n][index1][index2] = hx_a_sum_array[n][index1]*ha3_temp[index2][1];
+ d_x_d_ha2 = hx_a_sum_array[n][index1]*ha3_temp[index2][0];
+ d_y_d_ha2 = hx_a_sum_array[n][index1]*ha3_temp[index2][1];
}
}
float d_mean_d_ha2 = 0.0f;
float d_dev_d_ha2 = 0.0f;
- d_mean_d_ha2 = exp(hxhh_a_sum_array[n][0])/(1.0f+exp(hxhh_a_sum_array[n][0]))*d_x_d_ha2[n][index1][index2];
- d_dev_d_ha2 = exp(hxhh_a_sum_array[n][1])/(1.0f+exp(hxhh_a_sum_array[n][1]))*d_y_d_ha2[n][index1][index2];
+ d_mean_d_ha2 = exp(hxhh_a_sum_array[n][0])/(1.0f+exp(hxhh_a_sum_array[n][0]))*d_x_d_ha2;
+ d_dev_d_ha2 = exp(hxhh_a_sum_array[n][1])/(1.0f+exp(hxhh_a_sum_array[n][1]))*d_y_d_ha2;
G_ha2[index1][index2] = G_ha2[index1][index2] + advantage[n]/pi_old[n]*(d_mean_d_ha2*Grad_Normal_Dist_Mean(mean_array[n],deviation_array[n],action_array[n])+d_dev_d_ha2*Grad_Normal_Dist_Deviation(mean_array[n],deviation_array[n],action_array[n]));
}
@@ -698,18 +688,20 @@
}
for (int n=0; n<batch_size; n++) {
+ float d_x_d_ba2 = 0.0f;
+ float d_y_d_ba2 = 0.0f;
if((advantage[n] >= 0.0f && ratio[n] >= 1.0f + epsilon) || (advantage[n] < 0.0f && ratio[n] < 1.0f - epsilon)) {
G_ba2[index2] = G_ba2[index2];
} else {
if (hxh_a_sum_array[n][index2] >= 0) {
- d_x_d_ba2[n][index2] = ha3_temp[index2][0];
- d_y_d_ba2[n][index2] = ha3_temp[index2][1];
+ d_x_d_ba2 = ha3_temp[index2][0];
+ d_y_d_ba2 = ha3_temp[index2][1];
}
float d_mean_d_ba2= 0.0f;
float d_dev_d_ba2= 0.0f;
- d_mean_d_ba2 = exp(hxhh_a_sum_array[n][0])/(1.0f+exp(hxhh_a_sum_array[n][0]))*d_x_d_ba2[n][index2];
- d_dev_d_ba2 = exp(hxhh_a_sum_array[n][1])/(1.0f+exp(hxhh_a_sum_array[n][1]))*d_y_d_ba2[n][index2];
+ d_mean_d_ba2 = exp(hxhh_a_sum_array[n][0])/(1.0f+exp(hxhh_a_sum_array[n][0]))*d_x_d_ba2;
+ d_dev_d_ba2 = exp(hxhh_a_sum_array[n][1])/(1.0f+exp(hxhh_a_sum_array[n][1]))*d_y_d_ba2;
G_ba2[index2] = G_ba2[index2] + advantage[n]/pi_old[n]*(d_mean_d_ba2*Grad_Normal_Dist_Mean(mean_array[n],deviation_array[n],action_array[n])+d_dev_d_ba2*Grad_Normal_Dist_Deviation(mean_array[n],deviation_array[n],action_array[n]));
}
@@ -720,27 +712,25 @@
float G_ha3[num_hidden_unit2][2] = {0.0f};
float G_ba3[2] = {0.0f};
- float d_x_d_ha3[batch_size][num_hidden_unit2][2] = {0.0f};
- float d_x_d_ba3[batch_size][2] = {0.0f};
- float d_y_d_ha3[batch_size][num_hidden_unit2][2] = {0.0f};
- float d_y_d_ba3[batch_size][2] = {0.0f};
for (int index2 = 0; index2 < 2; index2++) {
for (int index1 = 0; index1 < num_hidden_unit2; index1++) {
for (int n=0; n<batch_size; n++) {
+ float d_x_d_ha3 = 0.0f;
+ float d_y_d_ha3 = 0.0f;
if((advantage[n] >= 0.0f && ratio[n] >= 1.0f + epsilon) || (advantage[n] < 0.0f && ratio[n] < 1.0f - epsilon)) {
G_ha3[index1][index2] = G_ha3[index1][index2];
} else {
if (hxh_a_sum_array[n][index1] >= 0) {
if (hx_a_sum_array[n][index1] > 0) {
- d_x_d_ha3[n][index1][index2] = hxh_a_sum_array[n][index1];
- d_y_d_ha3[n][index1][index2] = hxh_a_sum_array[n][index1];
+ d_x_d_ha3 = hxh_a_sum_array[n][index1];
+ d_y_d_ha3 = hxh_a_sum_array[n][index1];
}
}
float d_mean_d_ha3 = 0.0f;
float d_dev_d_ha3 = 0.0f;
- d_mean_d_ha3 = exp(hxhh_a_sum_array[n][0])/(1.0f+exp(hxhh_a_sum_array[n][0]))*d_x_d_ha3[n][index1][index2];
- d_dev_d_ha3 = exp(hxhh_a_sum_array[n][1])/(1.0f+exp(hxhh_a_sum_array[n][1]))*d_y_d_ha3[n][index1][index2];
+ d_mean_d_ha3 = exp(hxhh_a_sum_array[n][0])/(1.0f+exp(hxhh_a_sum_array[n][0]))*d_x_d_ha3;
+ d_dev_d_ha3 = exp(hxhh_a_sum_array[n][1])/(1.0f+exp(hxhh_a_sum_array[n][1]))*d_y_d_ha3;
G_ha3[index1][index2] = G_ha3[index1][index2] + advantage[n]/pi_old[n]*(d_mean_d_ha3*Grad_Normal_Dist_Mean(mean_array[n],deviation_array[n],action_array[n])+d_dev_d_ha3*Grad_Normal_Dist_Deviation(mean_array[n],deviation_array[n],action_array[n]));
}
@@ -750,17 +740,19 @@
}
for (int n=0; n<batch_size; n++) {
+ float d_x_d_ba3 = 0.0f;
+ float d_y_d_ba3 = 0.0f;
if((advantage[n] >= 0.0f && ratio[n] >= 1.0f + epsilon) || (advantage[n] < 0.0f && ratio[n] < 1.0f - epsilon)) {
G_ba3[index2] = G_ba3[index2];
} else {
- d_x_d_ba3[n][index2] = 1.0f;
- d_y_d_ba3[n][index2] = 1.0f;
+ d_x_d_ba3 = 1.0f;
+ d_y_d_ba3 = 1.0f;
float d_mean_d_ba3= 0.0f;
float d_dev_d_ba3= 0.0f;
- d_mean_d_ba3 = exp(hxhh_a_sum_array[n][0])/(1.0f+exp(hxhh_a_sum_array[n][0]))*d_x_d_ba3[n][index2];
- d_dev_d_ba3 = exp(hxhh_a_sum_array[n][1])/(1.0f+exp(hxhh_a_sum_array[n][1]))*d_y_d_ba3[n][index2];
+ d_mean_d_ba3 = exp(hxhh_a_sum_array[n][0])/(1.0f+exp(hxhh_a_sum_array[n][0]))*d_x_d_ba3;
+ d_dev_d_ba3 = exp(hxhh_a_sum_array[n][1])/(1.0f+exp(hxhh_a_sum_array[n][1]))*d_y_d_ba3;
G_ba3[index2] = G_ba3[index2] + advantage[n]/pi_old[n]*(d_mean_d_ba3*Grad_Normal_Dist_Mean(mean_array[n],deviation_array[n],action_array[n])+d_dev_d_ba3*Grad_Normal_Dist_Deviation(mean_array[n],deviation_array[n],action_array[n]));
}
@@ -768,7 +760,7 @@
G_ba3[index2] = -G_ba3[index2] / batch_size;
//ba3_temp[index2] = ba3_temp[index2] - gradient_rate * G_ba3[index2];
}
-
+
// Simultaneous Update
for (int index2 = 0; index2 < num_hidden_unit1; index2++) {
for (int index1 = 0; index1 < num_input_RL; index1++) {
@@ -788,155 +780,9 @@
}
ba3_temp[index2] = ba3_temp[index2] - gradient_rate * G_ba3[index2];
}
-
+
}
-///////////////////////////ReLU - Bad performance//////////////////////////////////
-//void update_Actor_Networks(float (*arr)[num_input_RL])
-//{
-// float gradient_rate = 0.001f; //-0.01f
-//
-// float G_ha1[num_input_RL][num_hidden_unit] = {0.0f};
-// float G_ba1[num_hidden_unit] = {0.0f};
-// for (int index2 = 0; index2 < num_hidden_unit; index2++) {
-// for (int index1 = 0; index1 < num_input_RL; index1++) {
-// for (int i=0; i<batch_size; i++) {
-// if((advantage[i] >= 0.0f && ratio[i] >= 1.0f + epsilon) || (advantage[i] < 0.0f && ratio[i] < 1.0f - epsilon)) {
-// G_ha1[index1][index2] = G_ha1[index1][index2];
-// } else {
-//
-// float hx_sum_total = 0.0f;
-// for(int m = 0; m < num_hidden_unit; m++) {
-// for(int n = 0; n < num_input_RL; n++) {
-// hx_sum_total = hx_sum_total + ha1_temp[n][m]*arr[i][n];
-// }
-// }
-// hx_sum_total = hx_sum_total + bc1_temp[index2];
-// float d_mean_d_ha1 = 0.0f;
-// float d_dev_d_ha1 = 0.0f;
-// if (hx_sum_total >=0) {
-// float hx_sum = 0.0f;
-// for(int j=0; j<num_input_RL; j++) {
-// hx_sum = hx_sum + ha1_temp[j][index2]*arr[i][j];
-// }
-// hx_sum = hx_sum + bc1_temp[index2];
-// if (hx_sum >= 0) {
-// d_mean_d_ha1 = ha2_temp[index2][0]*arr[i][index1];
-// d_dev_d_ha1 = ha2_temp[index2][1]*arr[i][index1];
-// } else {
-// d_mean_d_ha1 = 0.0f;
-// d_dev_d_ha1 = 0.0f;
-// }
-// } else {
-// d_mean_d_ha1 = 0.0f;
-// d_dev_d_ha1 = 0.0f;
-// }
-// G_ha1[index1][index2] = G_ha1[index1][index2] + advantage[i]/pi_old[i]*(d_mean_d_ha1*Grad_Normal_Dist_Mean(mean_array[i],deviation_array[i],action_array[i])+d_dev_d_ha1*Grad_Normal_Dist_Deviation(mean_array[i],deviation_array[i],action_array[i]));
-// }
-// }
-// G_ha1[index1][index2] = G_ha1[index1][index2] / batch_size;
-// ha1_temp[index1][index2] = ha1_temp[index1][index2] - gradient_rate * G_ha1[index1][index2];
-// }
-// for (int i=0; i<batch_size; i++) {
-// if((advantage[i] >= 0.0f && ratio[i] >= 1.0f + epsilon) || (advantage[i] < 0.0f && ratio[i] < 1.0f - epsilon)) {
-// G_ba1[index2] = G_ba1[index2];
-// } else {
-//
-// float hx_sum_total = 0.0f;
-// for(int m = 0; m < num_hidden_unit; m++) {
-// for(int n = 0; n < num_input_RL; n++) {
-// hx_sum_total = hx_sum_total + ha1_temp[n][m]*arr[i][n];
-// }
-// }
-// hx_sum_total = hx_sum_total + bc1_temp[index2];
-// float d_mean_d_ba1 = 0.0f;
-// float d_dev_d_ba1 = 0.0f;
-// if (hx_sum_total >=0) {
-//
-// float hx_sum = 0.0f;
-// for(int j=0; j<num_input_RL; j++) {
-// hx_sum = hx_sum + ha1_temp[j][index2]*arr[i][j];
-// }
-// hx_sum = hx_sum + bc1_temp[index2];
-//
-// if(hx_sum >=0) {
-// d_mean_d_ba1 = ha2_temp[index2][0];
-// d_dev_d_ba1 = ha2_temp[index2][1];
-// } else {
-// d_mean_d_ba1 = 0.0f;
-// d_dev_d_ba1 = 0.0f;
-// }
-// } else {
-// d_mean_d_ba1 = 0.0f;
-// d_dev_d_ba1 = 0.0f;
-// }
-// G_ba1[index2] = G_ba1[index2] + advantage[i]/pi_old[i]*(d_mean_d_ba1*Grad_Normal_Dist_Mean(mean_array[i],deviation_array[i],action_array[i])+d_dev_d_ba1*Grad_Normal_Dist_Deviation(mean_array[i],deviation_array[i],action_array[i]));
-// }
-// }
-// G_ba1[index2] = G_ba1[index2] / batch_size;
-// ba1_temp[index2] = ba1_temp[index2] - gradient_rate * G_ba1[index2];
-// }
-//
-// float G_ha2[num_hidden_unit][2] = {0.0f};
-// float G_ba2[2] = {0.0f};
-// for (int index2 = 0; index2 < 2; index2++) {
-// for (int index1 = 0; index1 < num_hidden_unit; index1++) {
-// for (int i=0; i<batch_size; i++) {
-// if((advantage[i] >= 0.0f && ratio[i] >= 1.0f + epsilon) || (advantage[i] < 0.0f && ratio[i] < 1.0f - epsilon)) {
-// G_ha2[index1][index2] = G_ha2[index1][index2];
-// } else {
-//
-// float hx_sum_total = 0.0f;
-// for(int m = 0; m < num_hidden_unit; m++) {
-// for(int n = 0; n < num_input_RL; n++) {
-// hx_sum_total = hx_sum_total + ha1_temp[n][m]*arr[i][n];
-// }
-// }
-// hx_sum_total = hx_sum_total + bc1_temp[index2];
-// float d_mean_d_ha2 = 0.0f;
-// float d_dev_d_ha2 = 0.0f;
-// if (hx_sum_total >=0) {
-// float hx_sum = 0.0f;
-// for(int j=0; j<num_input_RL; j++) {
-// hx_sum = hx_sum + ha1_temp[j][index1]*arr[i][j];
-// }
-// hx_sum = hx_sum + bc1_temp[index1];
-// if (hx_sum >= 0) {
-// d_mean_d_ha2 = hx_sum;
-// d_dev_d_ha2 = hx_sum;
-// } else {
-// d_mean_d_ha2 = 0.0f;
-// d_mean_d_ha2 = 0.0f;
-// }
-// } else {
-// d_mean_d_ha2 = 0.0f;
-// d_mean_d_ha2 = 0.0f;
-// }
-// G_ha2[index1][index2] = G_ha2[index1][index2] + advantage[i]/pi_old[i]*(d_mean_d_ha2*Grad_Normal_Dist_Mean(mean_array[i],deviation_array[i],action_array[i])+d_dev_d_ha2*Grad_Normal_Dist_Deviation(mean_array[i],deviation_array[i],action_array[i]));
-// }
-// }
-// G_ha2[index1][index2] = G_ha2[index1][index2] / batch_size;
-// ha2_temp[index1][index2] = ha2_temp[index1][index2] - gradient_rate * G_ha2[index1][index2];
-// }
-// for (int i=0; i<batch_size; i++) {
-// if((advantage[i] >= 0.0f && ratio[i] >= 1.0f + epsilon) || (advantage[i] < 0.0f && ratio[i] < 1.0f - epsilon)) {
-// G_ba2[index2] = G_ba2[index2];
-// } else {
-//
-// float d_mean_d_ba2 = 0.0f;
-// float d_dev_d_ba2 = 0.0f;
-// d_mean_d_ba2 = 1.0f;
-// d_dev_d_ba2 = 1.0f;
-// G_ba1[index2] = G_ba1[index2] + advantage[i]/pi_old[i]*(d_mean_d_ba2*Grad_Normal_Dist_Mean(mean_array[i],deviation_array[i],action_array[i])+d_dev_d_ba2*Grad_Normal_Dist_Deviation(mean_array[i],deviation_array[i],action_array[i]));
-// }
-// }
-// G_ba2[index2] = G_ba2[index2] / batch_size;
-// ba2_temp[index2] = ba2_temp[index2] - gradient_rate * G_ba2[index2];
-// }
-//}
-
-
-
float rand_normal(double mean, double stddev)
{
//Box muller method
@@ -1143,7 +989,7 @@
//read_field(i2c_slave_addr1);
//if(DIR_VALVE_ENC < 0) value = 1023 - value;
-// timer_while ++;
+ //timer_while ++;
///////////////////////////////////////////////////////Neural Network
@@ -1268,7 +1114,7 @@
hxh_c_sum_array[n][i] = hxh_c_sum[i];
}
hxhh_c_sum_array[n] = hxhh_c_sum;
-
+
pi[n] = exp(-(action_array[n]-mean_array[n])*(action_array[n]-mean_array[n])/(2.0f*deviation_array[n]*deviation_array[n]))/(sqrt(2.0f*PI)*deviation_array[n]);
Actor_Network_Old(temp_array);
pi_old[n] = exp(-(action_array[n]-mean_old)*(action_array[n]-mean_old)/(2.0f*deviation_old*deviation_old))/(sqrt(2.0f*PI)*deviation_old);
@@ -1305,7 +1151,6 @@
update_Critic_Networks(state_array);
update_Actor_Networks(state_array);
}
- //virt_pos = 10.0f;
Update_Done_Flag = 1;
Update_Case = 0;
//logging1 = V[0];